AI Cognition · LeverageAI · Cycle Compression

Cycle Compression: The Breathing Flywheel

AI didn’t just improve your thinking loop — it raised its frequency until the world answered while the thought was still warm. The ebook is the exhale that caches the upgrade.

Scott Farrell — LeverageAI · A second edition of the Learning Flywheel, Worldview Recursive Compression, and The Upgraded User · ~14 min read

Ten years ago I would have called a good month one or two worthwhile new ideas. Now, on a hot stretch, it feels more like ten to twenty in a day. That is not a lab measurement and I will not dress it up as one. It is a shape claim from the inside of a practice that finally has agents, tools, and a publishing stack that keep up with the brain. Even if the numbers are wrong by an order of magnitude, the phenomenon survives: something about the cycle time of cognition has changed.

Most AI writing still talks as if the win is quality. Better answers. Smarter drafts. Cleaner code. Those matter. They are also incomplete. What I am living is different:

Each cycle does not only make the next cycle better. Each cycle makes the next cycle happen sooner.

That is cycle compression. And it is the second edition of three frameworks I already published as full treatments: the AI Learning Flywheel, Worldview Recursive Compression, and The Upgraded User (public face: The Model Release That Upgraded My Brain).^1,2,3 Those pieces established that the loop exists, that worldview can be compiled, and that the third upgrade is the human. This piece names what they did not fully name: the frequency of the loop, the fractal breath that runs at every scale, and the cache role of the ebook.

What the parents already owned

The AI Learning Flywheel described co-evolution: better inputs produce better outputs, those outputs teach you, and the cycle escalates through Exposure, Critical Engagement, Self-Awareness, and Co-Evolution.¹ It is a quality story. It is true. It is not the whole story I am living now.

Worldview Recursive Compression made the loop architectural. World experience moves through you into explicit frameworks, into an operating system the models can load, into artefacts that expose gaps, and those gaps feed back into the frameworks. Frameworks are source code; outputs are regenerable binaries.² That is the formal compression ladder. What it did not emphasise is the tempo at which the ladder is now climbed.

The Upgraded User insists the model upgrade can land in the person, not only in the artefact. The conversation births ideas fast and fragile; the durable upgrade is what happens when the stretched version comes back and you actually re-enter it.³ That field guide owns elaborated rehearsal and read-back pedagogy. This article will not re-teach that. It will name the artefact that makes the landing possible: the ebook as augmentation checkpoint.

The enabling inequality

Classic psychology already knew that unrehearsed material in short-term memory does not sit around politely. In the Brown–Peterson tradition, recall collapses over seconds when rehearsal is blocked — often summarised as a roughly eighteen-second horizon for simple items under interference.^4,5 I am not claiming a multi-hour conceptual network decays on that clock. I am claiming the opposite structure of the old knowledge-work day:

THOUGHT DECAY TIME
        <
WORLD RESPONSE TIME

idea Monday
implement three weekends
result arrives
"oh yeah, that thing"

network is cold

WORLD RESPONSE TIME
        <
THOUGHT DECAY TIME

idea 09:00
evidence 11:30
"I'm still in this"

network is warm
→ association
→ next idea

That is the phrase I want you to keep:

AI has shortened world-response latency below the decay time of the original thought.

Previously the world answered after you had cognitively moved on. Now the answer collides with a live thought. Collision produces another thought. Agents amplify this because a long-running /goal can keep the execution side alive for hours while you dispatch six other thoughts; results return asynchronously into a day that is already a cognition event storm.

Industry is watching the same latency collapse from the software side. Anthropic’s 2026 agentic coding report describes agent-driven implementation, testing, and documentation collapsing cycle time from weeks to hours — tighter feedback loops, faster learning.⁶ That is not my idea-rate number. It is independent evidence that the world’s answer time for serious work has dropped by large factors. The personal claim is what happens when that drop crosses the half-life of the thought that launched the work.

On my own rough shape numbers — roughly 1–2 ideas a month then, 10–20 a day now on a hot stretch — people like to compute a 150× to 600× ratio. I will not defend the arithmetic. I will defend the regime change: the flywheel increased its own RPM. In conversation I have called that rise on the order of two orders of magnitude. Treat that as a shape claim, not a paper result.

Breathing is better than spinning

A flywheel suggests motion around the same axis. What I actually do is change resolution.

Inhale — descend into reality

A high-level thought — maybe old archives matter to the wiki — becomes code, databases, Lotus Notes DLLs, Wine, JSONL, Postgres, failures, weird edge cases. The agent goes down the rabbit hole for six hours. Something surprising happens. Reality returns evidence you could never have invented from the armchair.

Exhale — rise back into abstraction

Implementation detail becomes observation becomes pattern becomes the big coloured block becomes a framework becomes an ebook or a conversation like this. You name what changed in the worldview. That name becomes a new perceptual lens. The lens creates a new intent. The intent becomes a new goal. Back down.

// vertical cognition, not only circular
         ABSTRACT
            ▲
            │  ebook / framework
            │  exhale  — compress meaning
            │
            │  inhale  — expand into reality
            │  code / data / agents
            ▼
         CONCRETE

Stay high too long and you get nice theories with no contact. Stay low too long and you get POSTGRES! SQL! JSONL! forever. You need the respiration.

Ten years ago one full breath might take a month: idea, research week, three weekends of coding, maybe a write-up. Now a day can hold multiple complete respirations — shape a goal at 9:15, agent returns at noon, surprise by 12:20, new framework by 13:30, ebook underway by 14:00, gap exposed by 16:00, second agent descending. That is not “I type faster.” That is multiple cognitive respirations per day.

The pattern is fractal

I used to describe the breath only at framework scale. Then I watched an agent inside a ten-year-old archive: SQL probe, interpret, regex filter, interpret, forge a new instrument, interpret again. Same inhale/exhale. Every ten seconds. Inside the rabbit hole.

The respiration is scale-invariant. Abstraction descends until it hits reality. Reality returns resistance and evidence. Intelligence rises enough to reinterpret. Then it descends again with a better instrument.

At micro scale, deterministic code is not merely a “tool.” When the model faces thirteen gigabytes it cannot look at, and manufactures a SQL query that returns two hundred rows, then a filter that yields twelve patterns, it is growing a temporary sense organ. It forges an instrument for the question it has right now, measures, interprets, forges the next. Sometimes the instrument is disposable. Sometimes it becomes a C++ NSF driver nobody planned to productise. Either way: code reshapes the world into a representation intelligence can reason about.

// micro breath inside the hole
LLM: what is in here?
  ↓ SQL probe
RESULTS: mostly noise
  ↓ LLM: wrong shape
  ↓ grep / filter
NARROWER RESULTS
  ↓ LLM: wait — real staff discussions
  ↓ new probe
BETTER RESULTS
  ↓ LLM: that is the pattern

Pure LLM without contact is plausible talk. Pure determinism without interpretation is row dumps. The freakish capability is the ping-pong: AI narrows the question; determinism narrows the world; together they stay in contact with reality.

Code expands. The ebook compresses.

This is why the ebook is not “content marketing after the real work.”

Without the exhale you accumulate stimulation: result, idea, surprise, new project, more code. The ebook forces the rise. What was that actually about? What changed in the worldview? What is reusable? That is exactly the motion Worldview Recursive Compression formalised — experience through frameworks into artefacts that feed the next kernel.² The missing name was mechanical:

The ebook is the compression stroke of the cognitive engine.

Then the compressed framework becomes a new lens. You look back at reality through it. You see another thing to code. The breath continues.

// updated flywheel with respiration
INTENT
  ↓ delegate goal
DESCEND INTO REALITY
  ↓ code / data / agents
UNEXPECTED EVIDENCE
  ↓ rise / reflect
COMPRESS WORLDVIEW
  ↓ framework / ebook
NEW PERCEPTUAL LENS
  ↓ see new intent
  ↺  and each loop shortens the cost
      and time of the next loop

Better thinking creates better goals. Better goals create faster contact with reality. Faster contact returns evidence before the thought goes cold. Live thoughts collide with evidence and generate more ideas. Those ideas compress into frameworks that improve the next goal. The cycle accelerates itself.

The cache half: keep the augmentation

There is a second reason the ebook matters, and it is economic rather than poetic.

A thought starts as a line in a conversation. Then it runs through model elaboration, external research, competing concepts, structural pressure, writing judgement, and your existing canon. What you keep is not a transcript. It is a synthetically augmented thought. If you only leave it in the chat window, the model’s additional cognition disappears when context rolls away. If you compile it into an ebook, you cache the augmentation.

I have said this to myself in a line that stuck: the ebook was your first cognition wiki page. Not because it is a wiki. Because it is the first durable addressable surface where an upgraded thought sits still long enough for a later version of you to load it.

That is why I do not primarily read my own blog posts. I read the ebooks. The post is the public projection. The ebook is the artefact the author actually consumes. During a high-output window I can ship a burst of posts with linked deeper guides — often roughly three times the article — and the honest retrieval discipline is: post for the thesis, follow the deep link for the thinking. The Model Release piece already points at this: the idea is born fast and fragile in conversation; the upgrade wants a later, stretched form.³ How you rehearse that form is Upgraded User territory. That the form must exist as a cache is this article’s territory.

There is a failure mode hiding here. When the furnace is hot, you generate checkpoints faster than you load them into the human runtime. Multiple versions of you coexist in one week, each slightly more upgraded in the artefacts than in the wetware. Naming the cache does not automatically fill it back into the person — it only stops you pretending the chat log was enough.

The perturbation clause

Talking to a long-context model that shares hundreds of hours of vocabulary is not mainly a sycophancy problem. It is a path dependence problem. Shared context creates conceptual gravity. New observations fall toward existing concepts. That is useful: you get recognition instead of tourist-level paraphrase. It is also resonance. The closed Scott ↔ AI loop can become an echo chamber even when the model is adversarial, because the prior canon makes some frequencies louder.

Web research, in that light, is not a citations chore. It is an external perturbation source.

SCOTT ↔ AI
closed cognitive loop
        │
        │ can become resonant
        ▼
      WEB
alien language
existing literature
contradictory evidence
parallel ideas
        │
        ▼
loop deforms
thought breathes

Someone who has never heard of your framework kicks the thought sideways. Then your writing stack works the foreign material through your philosophy rather than pasting research on the end. That is why a good ebook can surprise its own author. It is re-processed cognition, not a transcript with footnotes.

The flywheel accumulates energy across cycles. The breath changes the volume of a thought within a cycle. Your system needs both. Perturbation keeps the breath from becoming a pure sine wave of self-agreement.

The publishing stack is a resolution ladder

Once you see the breath, the publishing architecture stops looking like marketing layers and starts looking like progressive resolution of the same cognition:

BLOG POST
  public projection / entry point / index
  enough to identify, address, and link

EBOOK / FIELD GUIDE
  the thinking artefact the author reads
  compression stroke + cache

WIKI
  compiled canonical world of concepts and edges

RAW CONVERSATION / CODE / EXPERIENCE
  bronze — source territory

The post is a semantic projection. It is not supposed to hold every bit of the source. It holds enough meaning to make the concept addressable and to point at the deeper object. The ebook is the deeper-resolution layer. The wiki is the collidable canon. The conversation is bronze.

When write latency falls below the lifetime of the conversation that generated the idea, something slightly absurd happens: you can be arguing idea B at 8 pm and search your own published history for idea A that you formulated at lunch. Your past is only a few hours behind you. That is not a content calendar. That is a cognition system whose output has become context for the next thought before the first thought cools.

What to protect

If you only take three operating rules from this second edition, take these:

1. Protect the inequality. Design work so world-response (agent runs, probes, experiments, drafts that hit reality) returns while you are still cognitively warm. Latency is not only a UX metric; it is a creativity constraint.
2. Run the breath at every scale. Macro: code and field contact paired with framework compression. Meso: goals that return evidence, not only status. Micro: LLM and deterministic instruments in ping-pong, not as rival religions.
3. Keep the compression stroke. An unread checkpoint is still better than an evaporated chat. The ebook (or your equivalent deep artefact) is how augmentation survives the context window. Loading it back into the human is the Upgraded User’s job; building and retaining it is yours here.

The 2025 flywheel said, in essence: better inputs → better outputs → better thinking → better inputs.¹ The 2026 lived version is becoming:

Better thinking creates better goals. Better goals create faster contact with reality. Faster contact returns evidence before the thought goes cold. Live thoughts collide with new evidence and generate more ideas. Those ideas compress into frameworks that improve the next goal. The cycle accelerates itself.

The flywheel developed respiration. Then AI raised the respiration rate until the world started answering inside the half-life of the thought. That is cycle compression. That is the breathing flywheel. And the artefact on the exhale — the one you keep — is how the augmentation stops being a mood and becomes a cache.